ForecastingForecasting
Discussion of forecasting methods, as well as specific forecasts relevant to doing good
Posts tagged Forecasting
Quick takes
Forecasting is an important tool for improving the future, because good forecasts and estimates can help us appropriately plan interventions and assess risks. Over the past several decades there has been significant research and investment in forecasting and estimation techniques, tools, and organizations. This continues to be an area of investment for improving our ability to make good decisions.
There are some major branches of forecasting within the EA movement:
Institutional forecasting.
Forecasting in institutions can range from predicting broad metrics to specific outcomes based on specific decisions. There can often be problems with buy-in from key stakeholders, who either see this as an unnecessary step or are concerned for their own status.
Forecasting is hard but many top forecasters use common techniques. This suggests that forecasting is a skill that can be learnt and practised.
Reference Class Forecasting on Wikipedia
Suppose we are trying to find the probability that an event will occur within the next 5 years. One good place to start is by asking "of all similar time periods, what fraction of the time does this event occur?". This is the base rate.
If we want to know the probability that Joe Biden is President of the United States on Nov. 1st, 2024, we could ask
These are all examples of using base rates. [These examples are taken from Base Rates and Reference Classes by jsteinhardt.]
Base rates represent the outside view for a given question. They are a good place to start but can often be improved on by updating the probability according to an inside view.
Note that there are often several reference classes we could use, each implying a different base rate. The problem of deciding which class to use is known as the reference class problem.
A forecaster is said to be calibrated if the events they say have a X% chance of happening, happen X% of the time.
Most people are overconfident. When they say an event has a 99% chance of happening, often the events happen much less frequently than that.
This natural overconfidence can be corrected with calibration training. In calibration training, you are asked to answer a set of factual questions, assigning a probability to each of your answers.
A list of calibration training exercises can be found here.
Much like Fermi estimation, questions about future events can often be decomposed into many different questions, these questions can be answered, and the answers to these questions can be used to reconstruct an answer to the original question.
Suppose you are interested in whether AI will cause a catastrophe by 2100. For AI to cause such an event, several things need to be true: (1) it needs to be possible to build advanced AI with agentic planning and strategic awareness by 2100, (2) there need to be strong incentives to apply such a system, (3) it needs to be difficult to align such a system should it be deployed, (4) a deployed and unaligned AI would act in unintended and high-impact power seeking ways causing trillions of dollars in damage, (5) of these consequences will result in the permanent disempowerment of all humanity and (6) this disempowerment will constitute an existential catastrophe. Taking the probabilities that Eli Lifland assigned to each question gives a 80%, 85%, 75%, 90%, 80% and 95% chance of events 1 through 6 respectively. Since each event is conditional on the ones before it, we can find the probability of the original question by multiplying all the probabilities together. This gives Eli Lifland a probability of existential risk from misaligned AI before 2100 to be approximately 35%. For more detail see Eli's original post here.
Decomposing questions into their constituent parts, assigning probabilities to these sub-questions, and combining these probabilities to answer the original questions is believed to improve forecasts. This is because, while each forecast is noisy, combining the estimates from many questions cancels the noise and leaves us with the signal.
Question decomposition is also good at increasing epistemic legibility. It helps forecasters to communicate to others why they've made the forecast that they did and it allows them to identify their specific points of disagreement.
A premortem is a strategy used once you've assigned a probability to an event. You ask yourself to imagine that the forecast was wrong and you then work backwards to determine what could potentially have caused this.
It is simply a way to reframe the question "in what ways might I be wrong?" but in a way that reduces motivated reasoning caused by attachment to the bottom line.
Getting Started on the Forecasting Wiki
While the above techniques are useful, they are no substitute for actually making predictions. Get out there and make predictions! Use the above techniques. Keep track of your predictions. Periodically evaluate questions that have been resolved and review your performance. Assess the degree to which you are calibrated. Look out for systematic mistakes that you might be making. Make more predictions! Over time, like with any skill, your ability can and should improve.
Other resources include:
For many years there have been calls to apply forecasting techniques to non-academic domains including journalism, policy, investing and business strategy. Several organisations now exist within these niche.
Metaculus is a popular and established web platform for forecasting. Their questions mainly focus on geopolitics, the coronavirus pandemic and topics of interest to Effective Altruism.
They host prediction competitions with real money prizes and collect and track public predictions made by various figures.
Cultivate Labs build tools that companies can use to crowdsource information from among their employees. This helps leadership to understand the consensus of people working on the ground and use this to improve the decisions they make.
Kalshi provide real money prediction markets on geopolitical events. The financial options they provide are intended to be used as hedges for political risk.
Manifold.Markets is a prediction market platform that uses play money. It is noteworthy for its ease of use, great UI and the fact that the market creator decides how the market resolves.
Fatebook is a website for rapidly recording personal predictions. Predictions are private by default, or can be shared with specific people or teams. You can use Fatebook on the website or via their integrations for Slack, Chrome extension (e.g. in Google Docs or Notion), or API integrations.
QURI is a research organisation that builds tools that make it easier to make good forecasts. Their most notable tool is Squiggle - a programming language designed to be used to make legible forecasts in a wide range of contexts.
This is a broad topic group that captures several sub-topics:
Help The Unjournal prioritize AI economics and governance research
(https://uj-prioritization-dashboard.netlify.app/ai-governance/)
The Unjournal (https://info.unjournal.org/) is prioritizing research on AI economics, governance, social and economic impacts, and risks from increasingly capable AI. (NB: We're not covering technical/CS/ML/AI-safety research.). We want to commission expert public evaluations and help synthesize, disseminate, and curate this work.
Much of this research seems 1) important/impactful/influential, (2) involves considerable 'firepower' from prominent researchers and practitioners, (3) not 'obviously true', and (4) time-sensitive: relevant for funding, policy, and research-steering now, but not after the "six months to six years" required for the traditional (economics) journal process
Our preliminary shortlist page (https://uj-prioritization-dashboard.netlify.app/ai-governance/) includes paper summaries, AI-generated audio walkthroughs, and an interface for giving feedback. We also invite suggestions fr other research (please check the broader list/interface to avoid duplication).
Especially looking for your input on:
Timing and next steps (our plan)
If it this works well, we hope to scale up and continue in the following months.
Alongside the evaluations, we want to help people find and use this research, and to help shape its direction. I hope the prioritization itself will be useful to students, researchers, practitioners, and other stakeholders. (NB: the "prioritization" here is mainly about "potential for impact and value for evaluation"; assessing the research credibility is for the evaluation part).
I wanted to share this quickly; planning a fuller post soon. (Feedback on this post itself is helpful too.)
Related: We’re also exploring sponsorship of evaluations/themes/questions (https://uj-prioritization-dashboard.netlify.app/sponsor/). Looking for feedback: Would this be useful, would you sponsor or recommend others to do so? Would sponsorship affect your trust in the evaluations (given the rules and safeguards mentioned in the link)?
I’m taking my second stab at formulating this AI bubble bet in a way that is attractive for people who might want to take the other side of it. I’m betting the AI bubble will pop within about 7 years. I’m looking for people who are willing to bet it won’t pop, or that there is no AI bubble.
I’m open to proposed modifications.
Stakes:
$20 to the charity of the winner’s choice.
Resolution criteria:
By January 1, 2034, all of the following news outlets will report — as fact, not opinion or editorial — that the AI bubble has popped:
To repeat, all of these news outlets have to report it. And they have to report it as news. Opinion columns and editorials don’t count.
Moreover, no major news publication classified as “generally reliable” by Wikipedia’s editors — either in general factual news information or in the area of finance, technology, economics, or markets — will report that the AI bubble has not actually popped. (Let’s say a “generally reliable” major news publication needs to disagree with the above listed news outlets within a month of all the listed outlets reporting that the bubble has popped.)
Moreover, if there is any survey of experts (e.g., economists, fund managers, institutional investors, financial analysts) on the matter by an organization generally considered reliable (e.g., Bank of America, which has conducted similar surveys in the past), at least 80% of respondents will agree the AI bubble as popped. If the survey finds that 79.9% or less of respondents agree, then I lose the bet. (Let's say this survey can happen within a month of when all the listed outlets report that the AI bubble has popped.)
Finally, all the following individual experts in finance or economics will agree that the AI bubble has popped:
Let’s say after all the above listed news outlets report that the AI bubble has popped, these individual experts have a month to publicly weigh in. If they express no opinion within a month, they’ll be removed from the list.
Explanation:
Whether the AI bubble has popped is difficult to operationalize and could conceivably involve tracking many different metrics, such as valuations, AI investment, investor sentiment, and so on. This is a complex job even for an expert. Deferring to qualified “judges” is much easier, and much easier to get right than attempting to operationalize it oneself.
The resolution criteria are designed to remove all doubt that, by some random fluke, a bunch of news outlets might report that the AI bubble has popped when it really hasn’t. If the AI bubble actually pops, it will probably be so clear and uncontroversial that someone will have no reservations about conceding it happened. (Conversely, if the AI bubble doesn’t pop, it will be equally obvious.)
In fact, the resolution criteria are so slanted toward the person taking the other side of that bet that if literally everyone in the world except Tyler Cowen agreed that the AI bubble had popped, I would still lose the bet.
The stakes are low because this bet is mainly just for fun and also to make a point.
The bet’s timeframe is about 7 years because I read that many AI investments are amortized over 5 to 7 years. It’s hard to know for sure what exactly AI investors’ time horizons might be, but 7 years is probably long enough. Even if the AI bubble were to pop in 8 years rather than 6 or 7, that would still be a different enough outcome than what I’m expecting that I would say I’ve lost the bet fair and square.
Again, I'm open to proposed modifications.
Humanity's inability to coordinate an AI slowdown may itself be early evidence that we are starting to lose control. Especially given all the recent omens around cyber and bio of frontier models, with open weight models just a few months away.
Will anyone take the other side of this bet? By January 1, 2034, the New York Times, Wall Street Journal, Financial Times, and Bloomberg will all agree the AI bubble has popped. (As in, they will all report it as news, not just publish an opinion column saying so.)
The stakes: $20 to the charity of the winner's choice.
Why this bet? Mainly just for fun. But also to make a point.
I chose 7 years as the time horizon for the bet because I read that 5 to 7 years is the time horizon for many AI investments.
My hunch is that the AI bubble will pop within 3 years, but who knows. Markets can keep lending and investing in unprofitable ventures for a very long time.
Preliminary quick take- Having 80% probablity of objective data might help outperform an entity with 74-75% accuracy with subjective data but it may just match an entity with objective data if there are no errors. If the accuracy and probablities remain stable.
It may help better to get probablistic data that is objective at a high enough rate that it outperforms others without knowing the type of data they have through higher accuracy by comparing their accuracy with one's own accuracy if one's accuracy is higher than others.
The steps could be-
Note- An AI was used to check whether this was good enough for a quick take before posting these steps and idea for preliminary quick take was given long time ago by AI excluding simultaneous discovery the idea was taken from me observing YouTube videos on getting full score on parallel based questions that seemed to have worked for me and that observation that the objective information seems to have beat(en) subjective information.
Additionally a mental model of a Jenga tower was created with bricks based on objective vs subjective information and whether the information was still valid with different block shapes. With the invalid shapes of blocks imagined to be taken out of that Jenga tower and assumed whether that tower falls for different information(created by ChatGPT although it might work the same in Gemini as I did not have time to test).
The idea was that both of them might have access to web information and other information potentially such as alternate data(probably) and they might give current information with the feelings that they might not give "stale data" with scary outcomes imagined in mind. With uncertanity("hallucinated" and imagined with a logo potentially protected by copyright appearing in the final image before my limit ran out as far as I can remember).
The observation might not be true always and I would say for caution as the image of the Jega tower might be protected by copyright but I do not know copyright law completely with complete understanding to say it. I had searched in Google "Are images by Jenga tower protected by copyright" and I got an AI's response that could be wrong and the things that could be used from here are the Comparative Asymetric Objective Probablity mental model and the Jenga tower mental model without breaking copyright and other laws.
Most AI-geopolitics analysis concentrates on frontier production. I've been exploring a complementary question: what happens to the capabilities the frontier leaves behind?
Downloadable models can persist after commercial support and meaningful developer control disappear. As inference and adaptation costs decline, some of these systems may become accessible to actors that could not have exploited them at release.
The longer piece develops this through the ideas of capability sedimentation, temporal capability arbitrage and “compute-poor but model-rich” states.
I'm especially interested in whether this should affect how the AI-governance community thinks about model-release decisions and proliferation below the frontier.
https://fabiofrettoli.substack.com/p/the-geopolitics-of-model-graveyards?r=sb1v&utm_campaign=post-expanded-share&utm_medium=web
I’m willing to bet that Anthropic’s revenue growth over the next year will be slower than its revenue growth over the last 3 years. I proposed a specific bet here. Anyone who wants can offer to take the other side of that bet. Or you can make a counteroffer.
I’m also willing to make a longer-term bet that the AI industry is in a bubble. I proposed a specific bet for that, too, here. Feel free to offer to take the other side of that bet or make a counteroffer.
I’d also be open to other bets. It seems pointless to bet about whether AGI or transformative AI will be deployed within the next 5-10 years, yet, for the heck of it, I would agree to a bet against that, too. (I’ll make bets for small, nominal amounts of money to be donated to the winner’s charity of choice, since the practical and legal problems with betting are too large otherwise.)
I’d also bet against the deployment of 100,000+ SAE Level 5 fully autonomous vehicles in North America within the next 3 years, if anyone has a strong opinion on that. I’d make a similar bet against the deployment of autonomous humanoid robots in North American households, although we’d have to come up with some specific resolution criteria.
Similarly, I’d bet against any significant level of near-term labour automation by LLMs or generative AI. Or against LLMs becoming capable of performing all sorts of specific tasks well.
On any of these topics, I’m also open to invitations for a public dialogue. (More on that topic here.)
Edit (2026-07-12 at 21:55 UTC): I’m still open to bets. If you’re interested, please contact me on Substack or by email.
Anthropic just released its S-1 form, which contains the official financial information it must report to the SEC in order to go public. From the S-1, we now know that Anthropic's revenue grew more than 100x (specifically, 112x) from 2023 to 2025. Revenue was:
A few people in the EA community apparently believe this revenue growth rate will continue for the next several years at least. However, I am confident we won't see Anthropic bring in $514 billion in revenue in 2027, which is what extrapolation would imply.
Indeed, I'm confident that Anthropic won't hit half of what extrapolation would imply, which is $257 billion in 2027.
I'm willing to bet as much and would be delighted to wager $20 on it, to be donated to the charity of the winner's choice.
I also have an open offer for a more general AI bubble bet.
Update (Friday, October 30th at 5:45pm Eastern):
Here's a good article on this topic. Two economists look at Anthropic's S-1 and ask whether a $2 trillion valuation is reasonable: https://econjared.substack.com/p/the-fantastical-promises-of-a-2-trillion